Scholarlink

This research addresses the critical challenge of patient data privacy in healthcare artificial intelligence by proposing and evaluating a federated learning framework for in-hospital mortality prediction. Leveraging neural networks, the extensive MIMIC-III dataset, and the IBM Granite foundational model, our approach enables collaborative model training across multiple institutions without centralized data aggregation, safeguarding sensitive patient information. The methodology details data preprocessing, the design of a decentralized federated architecture with a central aggregation server, and the configuration of a deep neural network model deployed via IBM WatsonX.ai. Utilizing IBM Granite improved computational efficiency by approximately 15%, accelerating model training and aggregation. Comparative analysis against a traditional centralized training model reveals that the federated approach achieves highly competitive predictive performance, with an AUC-ROC of 0.887 and an F1-Score of 0.628, closely matching the centralized model’s AUC-ROC of 0.895 and F1-Score of 0.644. These results underscore the viability of federated learning, augmented by IBM Granite and WatsonX.ai, as an effective and privacy-preserving paradigm for developing robust predictive models in clinical settings. This work contributes a solution that mitigates privacy concerns, fosters multi-institutional collaboration, and holds significant implications for improving patient outcomes through secure, accurate, and early risk assessment.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Decentralized Patient Mortality Prediction: A Federated Learning Approach with Neural Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Parth Bhatnagar,
  • Manjit Sodhi,
  • H. L. Gururaj,
  • M. Lavanya,
  • Bhargav P. Raj

摘要

This research addresses the critical challenge of patient data privacy in healthcare artificial intelligence by proposing and evaluating a federated learning framework for in-hospital mortality prediction. Leveraging neural networks, the extensive MIMIC-III dataset, and the IBM Granite foundational model, our approach enables collaborative model training across multiple institutions without centralized data aggregation, safeguarding sensitive patient information. The methodology details data preprocessing, the design of a decentralized federated architecture with a central aggregation server, and the configuration of a deep neural network model deployed via IBM WatsonX.ai. Utilizing IBM Granite improved computational efficiency by approximately 15%, accelerating model training and aggregation. Comparative analysis against a traditional centralized training model reveals that the federated approach achieves highly competitive predictive performance, with an AUC-ROC of 0.887 and an F1-Score of 0.628, closely matching the centralized model’s AUC-ROC of 0.895 and F1-Score of 0.644. These results underscore the viability of federated learning, augmented by IBM Granite and WatsonX.ai, as an effective and privacy-preserving paradigm for developing robust predictive models in clinical settings. This work contributes a solution that mitigates privacy concerns, fosters multi-institutional collaboration, and holds significant implications for improving patient outcomes through secure, accurate, and early risk assessment.